app.py

import gradio as gr
import torch
from PIL import Image
import os
from dotenv import load_dotenv
import spaces

load_dotenv()

# Disable torch compilation issues
torch._dynamo.config.disable = True
torch.backends.cudnn.allow_tf32 = True


IS_CUDA = torch.cuda.is_available()
IS_ZEROGPU = True if os.getenv("SPACES_ZERO_GPU", None) else False
if IS_ZEROGPU:
    torch.compiler.set_stance("force_eager")
    torch.set_float32_matmul_precision("high")
    torch.backends.cuda.matmul.allow_tf32 = True

MODEL_NAME  ="atlasia/AtlasOCR"
MAX_TOKENS = 4096


@spaces.GPU() 
@torch.inference_mode()
def predict(image: Image.Image) -> str:
    try:
        from unsloth import FastVisionModel
        model, processor = FastVisionModel.from_pretrained(
                    MODEL_NAME,
                    device_map="auto",
                    load_in_4bit=True,
                    use_gradient_checkpointing="unsloth",
                    token=os.environ["HF_API_KEY"],
                )
    except Exception as e:
        print(f"[Error] Failed to load model: {e}")
        raise  Exception(f"❌ Model failed to load: {e}")
        
    if image is None:
        gr.warning("Please upload an image.")

    # Build prompt
    messages = [
        {
            "role": "user",
            "content": [{"type": "image"}, {"type": "text", "text": ""}],
        }
    ]
    text = processor.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )

    inputs = processor(
        image,
        text,
        add_special_tokens=False,
        return_tensors="pt",
    ).to(model.device)

    with torch.no_grad():
        generated_ids = model.generate(
            **inputs,
            max_new_tokens=MAX_TOKENS,
            do_sample=False,
            temperature=0.0001, 
            pad_token_id=processor.tokenizer.eos_token_id,
        )

    # Trim input ids from output
    generated_ids = [
        out[len(inp) :] for inp, out in zip(inputs["input_ids"], generated_ids)
    ]
    text_out = processor.batch_decode(
        generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
    )
    return text_out[0].strip()




demo = gr.Interface(
    fn=predict,
    inputs=gr.Image(type="pil", label="Upload Image", height=400),
    outputs=gr.Textbox(
        label="Extracted Text",
        lines=20,
        show_copy_button=True,
        placeholder="Extracted text will appear here...",
    ),
    title="AtlasOCR - Darija Document OCR",
    description="Upload an image to extract Darija text.",
    examples=[["i3.png"], ["i6.png"]],
    )

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)